RLG-ALE/ALE family/Dataset 01

Agent's Last Exam

Gyms for GUI agents, built as an ALE-class dataset.

Dataset explorer

Inspect an RLG-ALE episode

Each task ships with a prompt, reference files, observations, and an evaluation record so you can see what the agent saw and why it passed.

Sample tasks / 05 included

Other task patterns

Create a price alertReact to a market postMove funds between instruments

Prompt

Add four representative tickers to the personal watchlist: a consumer staple, a manufacturer, a bank, and an insurer.

Reference files

{
  "id": "RLG-ALE-003",
  "family": "ALE",
  "environment": "android",
  "goal": "Add four representative tickers to the personal watchlist.",
  "constraints": ["Use search, not deep links", "Confirm each symbol is saved"]
}

Episode timeline

1Open the market appobservation

2Search for each requested assetobservation

3Add the asset to the watchlistobservation

Supported environments

Resettable, inspectable, agent-ready

RLG-ALE / v0.1

Android 14Touch + text inputDeterministic resetOffline replayScreenshot observations

Sample Gyms

Real-World Training Data

Test your agents with exclusive datasets for real-world tasks, then inspect a long-horizon episode in a controllable environment.

Airgnb

2K real house data · 10K video feeds

Navigable stay listings and walkthrough data.

Ginstagram

20K social posts · 100K real users

Social posts and audience analytics.

Linkedout

100K real profiles · 30K companies

Professional profiles and company records.

Gexpedia

10K flights · 10K hotels

Flights and hotel inventory for travel workflows.

Airgnb · example tasks

Booking House

Family Vacation Planner

Publish House

Business Trip Planner

Scenic Drive Itinerary

Verifier Coverage

Smart Validation for Complex Tasks

Combine data-aware and perception-driven validation layers to confirm successful task completion even in complex, multi-step scenarios.

Database Verifier

For example, when an Agent clicks the 'like' button on a post, a new record is created in the database table; the validator checks the table's data to determine whether the task has been completed.

Fig. 01

UI Verifier

Determine whether a task is completed by observing changes in the UI — for example, by using Android UI automator to output XML layout files, or by using CUA models such as UI-TARS or Gelato.

Fig. 02

On-Premise

On-Premise Deployment with Full Customization

Install on your own servers with air-gapped security. Modify benchmarks, create custom environments, and keep all data within your infrastructure.

Fig. 03

Air-gapped runtime

Deploy the container on isolated clusters with encrypted volume mounts and zero outbound traffic.

Customizable stacks

Swap benchmark suites, inject proprietary datasets, and wire your own validators without breaking the core framework.

Enterprise governance

Integrate with SSO, audit logging, and policy engines so every experiment is compliant by design.

Accelerate your RL training with GBOX

Contact Us